Essential genes are genes that critical for the survival of an organism. The prediction of essential genes in bacteria can provide targets for the design of novel antibiotic compounds or antimicrobial strategies. Here we propose a deep neural network (DNN) for predicting essential genes in microbes. Our DNN-based architecture called DeeplyEssential makes minimal assumptions about the input data (i.e., it only uses gene primary sequence and the corresponding protein sequence) to carry out the prediction, thus maximizing its practical application compared to existing predictors that require structural or topological features which might not be readily available. Our extensive experimental results show that DeeplyEssential outperforms existing classifiers that either employ down-sampling to balance the training set or use clustering to exclude multiple copies of orthologous genes. We also expose and study a hidden performance bias that affected previous classifiers.The code of DeeplyEssential is freely available at https://github.com/ucrbioinfo/DeeplyEssential 1 Introduction 1 Essential genes are those genes that are critical for the survival and reproduction of an 2 organism [17]. Since the disruption of essential genes induces the death of an organism, 3 the identification of essential genes can provide targets for new antimicrobial/antibiotic 4 drugs [7, 13]. The set of essential genes is also critical for the creation of artificial 5 self-sustainable living cells with a minimal genome [16]. Essential genes have also been a 6 cornerstone in understanding the origin and evolution of organisms [18]. 7 The identification of essential genes via wet-lab experiments is labor intensive, 8 expensive and time consuming. Such experimental procedures include single gene 9 knock-out [3, 12], RNA interference and transposon mutagenesis [8, 32]. Moreover, these 10 experimental approaches can produce contradicting results [23]. With the recent 11 advances in high-throughput sequencing technology, computational methods for 12 predicting essential genes has become a reality. Some of the early prediction methods 13 used comparative approaches by homology mapping, see, e.g., [27, 43]. With the 14 introduction of large gene database such as DEG, CEG and OGEE [4, 25, 40], researchers 15 designed more complex prediction models using a wider set of features. These features 16 can be broadly categorized into (i) sequence features, i.e., codon frequency, GC content, 17 gene length [29, 35, 42], (ii) topological features, i.e., degree centrality, cluster 18 coefficient [1, 6, 24, 31], and (iii) functional features, i.e., homology, gene expression 19cellular localization, functional domain and molecular properties [5,9,23,30,39].Sequence based features can be directly obtained from the primary DNA sequence of 21 a gene and its corresponding protein sequence. Functional features such as network 22 topology requires knowledge of protein-protein interaction network, e.g., STRING and 23 HumanNET [15,37]. Gene expression and functional dom...